英伟达将不再向中国出售H200芯片,执行总裁Colette Kress周二表示,此举结束了该公司在美国禁止向中国出售先进芯片的争议性漏洞中的参与。
在首席执行官兼联合创始人黄仁勋上周离职后,公司首次财报电话会议上,中国获取领先AI加速器的能力成为一个重要话题。这一披露与英伟达宣布成立董事会级别委员会监督国家安全政策、出口管制合规性和全球贸易风险的同一天。
英伟达向中国销售H200芯片的交易由黄仁勋和总统唐纳德·特朗普促成。该安排要求芯片通过美国运输,以便政府收取25%的关税。
这条路径激怒了那些希望剥夺中国实验室计算能力并维持美国在AI领域领先地位的强硬派人士,但黄仁勋长期以来一直认为,美国在中国市场竞争更为重要。今年4月,他告诉播客主持人Dwarkesh Patel:“如果让美国科技产业放弃这个市场,那是对我们国家的损害。”
黄仁勋被罢免——部分原因是他与一家涉嫌走私芯片进入中国的公司CEO有染——恰逢周一发布的新规,禁止为远程服务中国客户的外国数据中心使用先进的美国芯片。
投资银行Jefferies预测,这些规定可能会使英伟达明年的数据中心收入减少440亿美元至620亿美元。Kress引用新的出口管制措施称,对Vera Rubin等先进芯片的需求积压意味着这些规定不会对公司的总收入产生影响。
从长远来看,公司可能会失去在新加坡、马来西亚、泰国和印度尼西亚的潜在客户。批评者声称,英伟达对中国依赖程度高达其收入的20%——其新加坡子公司的年收入在2025年达到1040亿美元,之后英伟达停止披露这些数据——但这显然会发生变化。
即使在英伟达决定停止发货H200之前,中国也基本拒绝接受这些芯片的交付。根据乔治城大学安全与新兴技术中心研究员Jacob Feldgoise的说法,中国政府更倾向于为本国芯片行业创造需求,即使主要的AI计算用户仍然希望使用性能更优的英伟达设备。
与此同时,中国的前沿实验室正在寻找其他方式获取英伟达芯片。所有中国领先的AI模型都显示出曾使用被禁止出口到中国的Nvidia Grace Blackwell芯片进行训练的证据。
这种训练可能是在中国境外的数据中心进行的,但根据新规,向这些设施的销售将被禁止。出口管制律师Dan Fisher-Owens告诉TechCrunch,如果大型适合训练AI模型的数据中心想要购买英伟达芯片,它们必须证明不服务于任何中国客户。
如果无法提供这种认证,新规可能会迫使它们转向其他芯片选项,并进一步侵蚀英伟达的市场份额。
当您通过文章中的链接购买时,我们可能会获得少量佣金。这不会影响我们的编辑独立性。
Tim Fernholz是一位撰写技术、金融和公共政策的记者。他密切关注私人航天工业的崛起,并著有《火箭亿万富翁:埃隆·马斯克、杰夫·贝佐斯与新太空竞赛》一书。此前,他在Quartz(全球商业新闻网站)担任高级记者超过十年,并在华盛顿特区开始了他的政治记者生涯。您可以通过电子邮件联系或验证来自Tim的联系信息,邮箱地址为[
[email protected]
(/cdn-cgi/l/email-protection),或通过加密消息发送至Signal上的tim_fernholz.21。
查看简介
查看原文
Nvidia will no longer sell its H200 chips in China, acting CEO Colette Kress said Tuesday, ending the company’s participation in a controversial loophole in the U.S. ban on the sale of advanced chips to the country.
China’s access to leading AI accelerators was an important topic during the company’s first earnings call since CEO and co-founder Jensen Huang stepped down last week. The disclosure comes the same day Nvidia announced it was creating a board-level committee to oversee national security policies, export control compliance, and global trade risks.
Nvidia’s deal to sell H200 chips in China was brokered by Huang and President Donald Trump. The arrangement required the chips to be shipped through the U.S., allowing the government to collect a 25% tariff.
The pathway angered China hawks who wanted to deny Chinese labs computing power and maintain the United States’ lead in AI, but Huang long argued that it was more important for the U.S. to compete in China. In April, he told podcaster Dwarkesh Patel that “to concede that market for the United States technology industry is a disservice to our country.”
Huang’s ouster — in part over an affair with the CEO of a company under investigation for smuggling chips into China — has coincided with new rules, issued Monday, barring advanced U.S. chips from foreign data centers that serve Chinese customers remotely.
Investment bank Jefferies forecast that those rules might trim between $44 billion and $62 billion from Nvidia’s data center revenue next year. Referencing the new export controls, Kress said that the backlog of demand for advanced chips like Vera Rubin means the rules would have no impact on the company’s top-line revenue.
In the long term, the company may face the loss of potential customers in Singapore, Malaysia, Thailand, and Indonesia. Critics of Nvidia’s dependence on China claim that as much as 20% of its revenue comes from demand linked to the country — annual revenue at its Singapore subsidiary hit $104 billion in 2025 before Nvidia stopped disclosing those figures — but that’s clearly going to change.
Even before Nvidia decided to stop shipping H200s, China had largely refused to accept their delivery. The government would prefer to generate demand for the country’s domestic chip industry, even if major AI compute users want the still-superior Nvidia equipment, according to Jacob Feldgoise, a researcher at Georgetown’s Center for Security and Emerging Technology.
Meanwhile,Chinese frontier labs are finding ways to access Nvidia chips elsewhere. All of China’s leading AI models show evidence of having been trained on Nvidia Grace Blackwell chips that are prohibited from being exported to China.
That training may have taken place at data centers outside of China, but sales to those facilities will be barred under the new rules. If large data centers suited for training AI models want to buy Nvidia chips, they will have to certify they aren’t serving any Chinese customers, export control attorney Dan Fisher-Owens told TechCrunch.
If they can’t provide that certification, the new rules may force them toward other chip options and further eat into Nvidia’s market share.
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.
Tim Fernholz is a journalist who writes about technology, finance and public policy. He has closely covered the rise of the private space industry and is the author of Rocket Billionaires: Elon Musk, Jeff Bezos and the New Space Race. Formerly, he was a senior reporter at Quartz, the global business news site, for more than a decade, and began his career as a political reporter in Washington, D.C. You can contact or verify outreach from Tim by emailing [email protected] or via an encrypted message to tim_fernholz.21 on Signal.
View Bio
原文版权归 TechCrunch · AI 及原作者所有,本站转载仅供阅读。如需删除,请联系 support@aisa.one。